Download Drew Kristensen, Dr. Adam A. Smith (Advisor) Department of

Drew Kristensen, Dr. Adam A. Smith (Advisor)
Department of Mathematics and Computer Science, University of Puget Sound
Introduction
The ultimate goal of this study is to gain a better understanding of
the affective state of mice, by analyzing mouse vocalizations. To
achieve this goal, this study aimed to distinguish vocalizations from
recordings using an artificial neural network. In order to analyze
these calls, they must first be identified and characterized. While
human technicians are very accurate at these classifications,
manually annotating the ultrasonic vocalizations (USVs) is a
time-intensive process and might be better if it were automated.
This project relies on neural networks, a machine learning
technique that employs weighted inputs with the ability to learn as a
method of discerning calls from noise.
Preliminary Results
Weights are represented
with black being the
highest value and white
being the lowest
From the raw spectrogram
(top) and the annotated
spectrogram (second from
top), the network multiplies
each time-value pair and
the surrounding 7×7 grid by
the 5 weight matrices (left)
and outputs the resulting
image marking all probable
call locations (below)
Background
A neural network consists of a number of artificial neurons which
take weighted inputs in order to calculate a value. Here, our
weighted inputs are the weights stored within our neurons, and the
time-frequency values found in the spectrogram. In order to do this,
we classify every time-frequency value as one of 6 categories;
rising, flat, and falling for calls, streak and spike noise, or neither
noise nor a vocalization. While a single neuron is capable of only
the most simple distinctions, with more neurons, the more
complicated the “learning” function can be. By sending the input
data to a layer of these neurons, we create a neural network. The
output of these neurons is taken as the input of a final neuron, so the
weights of the various neurons can be changed to better fit the data.
This results in the final classification of the input as “call” or
“no-call”. Because the neurons take in many weighted values and
calculate a sum on which to base the output on, the various weights
for each input value can be changed with each iteration. This
re-evaluation serves the purpose of ensuring that the network can
accurately converge upon the best values when working through the
trial set.
The artificial neural network (ANN) works by taking a
spectrogram generated from a preprocessed audio file. From this
spectrogram, a user must first mark the image in locations where
either calls or noise are present. The ANN then goes through the
spectrogram, time-frequency pair by time-frequency pair, multiplying
each value and the surrounding 7×7 grid of these values by a series of
weights stored in various matrices within a layer of neurons. If the
result of the operation is above a certain threshold, the ANN outputs a
positive value for a call. These neurons which hold these weighted
matrices can be trained if given an annotated version of a
spectrogram in order to better represent the data. If the value output
was too small and therefore below the threshold, the difference
between the target value and estimated value is passed through the
input matrices, allowing each value within them to be altered by an
amount relative to their overall weight to the final output. This
process also works for values which were higher than the actual
values, and thus need training down to lower values. Training allows
the ANN to key in on specific aspects of calls that may or may not be
apparent to human observers, yet work with high precision.
Motivation
Mice are a fundamental part of research in many scientific
fields because we are able to tailor them to fit our needs. These
animals provide a great testing medium for a wide range of
treatments which need experimentation or discovery. This is
possible since mice share many important anatomical and
physiological features with humans. Furthermore, we can alter
mice to fit our specific needs through careful use of breeding,
surgical changes, or alterations to their genetic makeup. These
modifications allow us to model specific diseases such as
cancer or mental disorders such as autism and schizophrenia.
This study’s aim is to create an additional tool to aid the
current analysis of a mouse’s affective state done through
traditional methods, in order to gain an insight into mice that
model these mental disorders.
References
G. P. Lahvis, E. Alleva, and M. L. Scattoni, “Translating Mouse Vocalizations: Prosody
and Frequency Modulation,” Genes Brain Behav, vol. 10, no. 1, pp. 4–16, Feb. 2011.
M. Mitchell, An introduction to genetic algorithms. Cambridge, Mass.: MIT Press,
1996.
Mitchell, Tom M. "Artificial Neural Networks." Machine Learning. New York:
McGraw-Hill, 1997. 81-127. Print.
Smith, Adam A. "Hidden Markov Models and Mouse Ultrasonic Vocalizations." XRDS:
Crossroads, The ACM Magazine for Students XRDS 21.4 (2015): 48-53. Web.
I would like to extend my sincerest gratitude to the Sherman
Fairchild Research Grant for making this research project possible,
and to Dr. Adam A. Smith, whose patience and support was much
appreciated.